Your browser is not supported. Please update it.

12 July 2006

Intel shooting for multi-core chip dominance by 2010

Tom’s Hardware Guide, the quaint US PC help site, turned technology database, laid out this week Intel’s future thinking in how it will deliver raw chip processing power going forwards and overtake the aggressive Sun Microsystems and IBM roadmaps.

The details come from internal Intel documents and are in stark contrast to the futures that Intel has painted in the past. Once it kept faith with Moores Law on a single cpu, now it is saying that Moores Law (double the processing power every 18 months) will be achieved using up to 32 processors on a single chip.

Behind all of this is a set of bad decisions that almost cost Intel the leadership in CPU chip design. It was only a few short years ago that Intel conceded that IBM’s 4 way processing cores were killing its single chip design route in the server markets, and it set about designing its first dual core cpu which is in the heart of new PCs today.

Intel is aiming to trump the performance of the Sun Niagara III chip, due to ship in mid 2009, with its own 32 processor chip, supporting 128 threads, out in early 2010, with each core running at only two thirds of the clock speed of today’s fastest Xeon CPU, but delivering fifteen times the performance.

Already Texas Instruments is offering the mobile phone world its OMAP 3 designs for incorporation into handsets in volume by 2008,which has four separate ARM processing cores, each allocated for different roles. Smugly at its announcement early this year TI said, ‘That’s right we’ve been doing multi-core for years, unlike Intel.’

But now that Intel has bought into the concept, it is planning to drive much of its research into parallelism and generating multiple core devices without massive loss of power used up in management tasks.

Given that the Cell processor is already sampling, and has a single controlling P5 PowerPC core and 8 slave parallel instruction engines, with target pricing low enough to include in cheap consumer electronic devices, IBM is already pretty much already heading down the same path as Intel, if not a little way ahead.

Intel’s first multi-core project is referred internally as “Kevet” and the final 32 core chip is talked about as Keifer (surely Keifer Sutherland is in 24, not 32).

The first Keifer chip is to be built in 32 nanometer geometry in eight processing clusters with four cores each. Each node will have direct access to 3 MB cache, with 8 such caches connected in a ring architecture.

Intel appears to believe that Sun’s Niagara is the danger to its existing cpu leadership and expects Niagara II, in 65 nm in 2007 with 8 cores. But AMD and IBM are also heading down this route and Intel, as the sudden and recent convert, clearly feels the need to overtake all three organizations by 2010.

It looks to us that an awful lot of processing can be done by chips that have that much power, many of them driving dramatic changes in High Definition TV capability, and set tops, as much as pushing up the power of desktop PCs and servers.